[alpsaur]

Scope

Reliability hardening for this best-effort fork, preparing v0.5.9:

• Preserve coordinate-manager topology in deterministic convolution; cache canonical maps and select a per-call non-fused backend.
• Fix generated-coordinate publication races, empty-map reference lifetimes, and zero-sized CUDA launches.
• Embed/check torch/CUDA/ABI build facts; pin binary wheels to their build-time torch minor.
• Validate installed wheels outside the checkout; keep GPU CI trusted/manual and releases draft-only until all wheel builds pass.
• Repair public gradcheck and analytic/kernel-map tests; make long real-data benchmarks opt-in.
• Add a fixed-input cold/warm U-Net benchmark and remove interpolation host scalar synchronization, with a full-neighborhood metadata fast path.

No speculative kernel rewrite, FP8, full torch.compile integration, or coordinate-manager redesign. Determinism is a same-device/build/settings convolution-forward contract, not bitwise deterministic training.

Local validation

Python 3.12, torch 2.9.1+cu128, CUDA 12.8, RTX 5090:

• Source regression suite: 142 passed, 15 skipped, 102 subtests passed.
• Isolated installed CUDA wheel, copied tests (not package): 142 passed, 15 skipped, 102 subtests passed; CPU and CUDA smoke passed.
• Ten independent determinism/generated-map runs: 12 passed each, including 40 generated-map trials per run.
• CPU-only wheel built from the pre-repair sdist (native code unchanged by the repairs): 90 passed, 64 skipped in isolated installed-wheel tests.
• Sdist includes required CUDA sources/backend stub and excludes generated metadata/binaries.
• Ruff correctness checks, diff hygiene, actionlint (custom gpu runner label allowed), and Sphinx HTML build pass. Sphinx retains existing documentation warnings.

GPU suite skips are 12 explicit opt-in real-data/performance jobs and three pre-existing legacy API/raw-gradcheck tests. CPU skips additionally cover GPU-only tests. These are not counted as passes.

Merge/release gates

• Independent whole-diff review plus native and packaging/CI scouts reconciled. Two findings repaired in 90f1735; each raising reviewer independently confirmed the fix. All 49 changed files covered.
• GitHub CPU matrix passes at final head 90f1735: all 8 jobs.
• CUDA and CPU sdist builds validated; after the two Python-only repairs, regenerated sdist bytes match the final sources, and the rebuilt installed CUDA wheel passes the expanded regression suite.
• v0.5.9 published: all four release wheels passed CI smoke, asset digests/metadata verified, and the actual downloaded cp312 wheel passed CPU/CUDA smoke plus 142 tests and 102 subtests on RTX 5090 (15 documented skips).

Candidate committed as alpsaur. Independent review, CI, and release-artifact gates complete; PR merged by alpsaur and v0.5.9 published.